Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Tree reduction and search depth question

Author: José Carlos

Date: 02:56:14 08/10/01

Go up one level in this thread


On August 10, 2001 at 05:37:06, Andrew Williams wrote:

>On August 10, 2001 at 04:30:20, José Carlos wrote:
>
>>  After completely rewriting my chess program (Averno), I've run into the
>>null-move world, which I haddn't in old versions.
>>  In old Averno I had:
>>
>>  -Reasonably good move ordering (hash move; good captures; killers; ...)
>>  -Transposition table
>>  -2 killers
>>  -Lazy eval
>>  -Check and recapture extension
>>  -Some out-of-check stuff in qsearch
>>
>>  With that, I was searching 6-7 plies in blitz (in midgame) and 7-9 in
>>standard. [Athlon 550]
>>
>>  I've read some open source programs to get a clue on how to search deeper, and
>>now I have:
>>
>>  -Nullmove (tried R=3 and R=2, not yet R=2/3)
>>  -Futility pruning
>>  -Same move ordering as before
>>  -Same hashing scheme
>>  -Same 2 killers
>>  -Same Lazy eval
>>  -Check and null-threat extension
>>  -No out-of-check stuff in qsearch
>>
>>  And now, I search 7-8 plies in blitz and 8-10 in standard (and the worst thing
>>is that the program looks weaker than the old one!)
>>  That looks like a poor search depth, isn't it? So what am I missing to get, as
>>most null-movers, 12-13 in standard?
>>  SEE? History heuristic for move ordering? ETC? Razoring?
>>  Sorry for _answering_ instead of _trying_ but, as I have so little free time
>>to work on my program, any advice would be appreciated.
>>  As well, I'd like to know if, with the stuff I've implemented so far, I should
>>expect a better performance. In that case, I might have a bug somewhere.
>>
>>  Thanks in advance,
>>
>>  José C.
>
>Seriously, I think the first thing you should look at is move-ordering. I find
>that errors I introduce here have an enormous impact on the depth of searches.
>One metric that many people use is to see what proportion of beta cut-offs in
>the search are achieved on the *first* move searched at a given node. The
>higher this figure the better. Aim for over 90%.
>
>Cheers
>
>Andrew

  Ok, I'll calculate that number tonight...

  José C.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.